BeCAPTCHA-Mouse: Synthetic Mouse Trajectories

and Improved Bot Detection

Alejandro Acien∗

, Aythami Morales, Julian Fierrez, Ruben Vera-Rodriguez

Biometrics and Data Pattern Analytics Lab, Universidad Autonoma de Madrid, Spain

Abstract

We first study the suitability of behavioral biometrics to distinguish between

computers and humans, commonly named as bot detection. We then present

BeCAPTCHA-Mouse, a bot detector based on neuromotor modeling of mouse

dynamics that enhances traditional CAPTCHA methods. Our proposed bot de-

tector is trained using both human and bot data generated by two new methods

developed for generating realistic synthetic mouse trajectories: i) a knowledge-

based method based on heuristic functions, and ii) a data-driven method based

on Generative Adversarial Networks (GANs) in which a Generator synthesizes

human-like trajectories from a Gaussian noise input. Experiments are conducted

on a new testbed also introduced here and available in GitHub: BeCAPTCHA-

Mouse Benchmark; useful for research in bot detection and other mouse-based

HCI applications. Our benchmark data consists of 10,000 mouse trajectories in-

cluding real data from 58 users and bot data with various levels of realism. Our

experiments show that BeCAPTCHA-Mouse is able to detect bot trajectories

of high realism with 93% of accuracy in average using only one mouse trajec-

tory. When our approach is fused with state-of-the-art mouse dynamic features,

the bot detection accuracy increases relatively by more than 36%, proving that

mouse-based bot detection is a fast, easy, and reliable tool to complement tra-

ditional CAPTCHA systems.

∗Corresponding author

Email addresses: alejandro.acien@uam.es (Alejandro Acien), aythami.morales@uam.es

(Aythami Morales), julian.fierrez@uam.es (Julian Fierrez), ruben.vera@uam.es (Ruben

Vera-Rodriguez)

Preprint submitted to Journal May 5, 2020

arXiv:2005.00890v1 [cs.CV] 2 May 2020

Keywords:

CAPTCHA, bot detection, behavior, biometrics, mouse, neuromotor

1. Introduction

How to distinguish between human users and artificial intelligence during

computer interactions is not a trivial task. This challenge was firstly discussed

by Alan Turing in 1950. He investigated whether machines could show an

intelligent behavior, and also how humans could be aware of these artificial be-

haviors. For this, he developed the famous Turing Test [1], commonly named as

The Imitation Game, in which a human evaluator would judge natural language

conversations between a human and a computer designed to generate human-

like responses. The Turing Test was both influential and widely criticized and

became an important concept in the artificial intelligence field [2]. However, at

the epoch of Alan Turing research, the problem of machines acting like humans

were commonly associated to science-fiction topics [3].

Nowadays, boosted by the last advances of machine learning technologies

and worldwide connections, that science-fiction topic becomes a real hazard. As

an example, bots are expected to be responsible for more than 40% of the web

traffic with more than 43% of all login attempts to come from malicious botnets

in the next years1. Malicious bots cause billionaire loses through web scraping,

account takeover, account creation, credit card fraud, denial of service attacks,

denial of inventory, and many other. Moreover, bots are used to influence and

divide society (e.g. usage of bots to interfere during Brexit voting day [4], or to

spread anxiety and sadness during the COVID-19 outbreak2,3through Twitter).

Bots are becoming more and more sophisticated, being able to mimic human

1https://resources.distilnetworks.com/white-paper-reports/bad-bot-report-2019

2https://www.washingtonpost.com/science/2020/03/17/analysis-millions-coronavirus-

tweets-shows-whole-world-is-sad/

3https://www.sciencealert.com/bots-are-causing-anxiety-by-spreading-coronavirus-

misinformation

online behaviors. On the other hand, algorithms to distinguish between humans

and bots are also getting very complex. We can distinguish two types of bot

detection methods in response to those sophisticated bots:

•Active Detection. Traditionally named as CAPTCHA (Completely Au-

tomated Public Turing test to tell Computers and Humans Apart), these

algorithms determinate whether or not the user is human by perform-

ing online tasks that are difcult for software bots to solve while being

easy for legitimate human users to complete. Some of the most popular

CAPTCHA systems are based on: characters recognition from distorted

images (text-based), class-objects identification in a set of images (image-

based), and speech translation from distorted audios (audio-based).

•Passive Detection. These detectors are transparent and analyze the users

behavior while they interact with the device. The last version of Google

reCAPTCHA v3 replaces traditional cognitive tasks by a transparent al-

gorithm capable of detecting bots and humans from their web behavior4.

Other researchers [5], describe browsing behavior of web users for detec-

tion of DDoS Attacks (Distributed Denial of Service).

Although these algorithms are broadly used, they present limitations. First

of all, ensuring a very accurate bot detection makes the tasks difficult to perform

even for humans. Second, most of the CAPTCHA systems can be easily solved

by the most modern machine learning techniques. For example, the text-based

CAPTCHA was defeated by Bursztein et al. [6] with 98% accuracy using a ML-

based system to segment and recognize the text. In [7], the authors designed an

AI-based system called unCAPTCHA to break Googles most challenging audio

reCAPTCHAs. Third, these algorithms process sensitive information and there

are important concerns about how they comply with new regulations such as the

European GDPR5. Fourth, the CAPTCHA systems become a great barrier to

4https://www.google.com/recaptcha/intro/v3.html

5https://complianz.io/google-recaptcha-and-the-gdpr-a-possible-conflict/

Classifier

Human

Bot

Human Mouse

Trajectories DB

Random

seed

Generator 1

Generator 2

Synthetic Mouse

Trajectories DB

Neuromotor

Model

Neuromotor

Model

Figure 1: Architecture of BeCAPTCHA-Mouse: Neuromotor features are extracted from

Human and Synthetic Mouse Tra jectories. A Classifier is then trained for bot detection. The

proposed Generators can be also helpful for other HCI applications.

people with visual or other impairments. Finally, the Turing Test was designed

as a task in which machines had to prove they were human, meanwhile in current

CAPTCHA systems humans have to prove they are not machines (e.g. Im not

a robot from Googles). This means that the responsibility to prove the users

humanity falls over human users instead of bots. At this point, there is still a

large room for improvement towards reliable bot detection able to stop malicious

software not bothering human users during natural web browsing.

On the other hand, biometric recognition refers to the automated recognition

of individuals based on their physiological (e.g. fingerprint, face,) and behavioral

(e.g. keystroke, gait) characteristics [8]. Traditionally focused on person recog-

nition, the individual patterns obtained from biometric signals characterize the

human being. Behavioral biometrics refers to those traits revealing distinctive

user behaviors and mannerisms when they interact with devices [9]. Behavioral

biometrics characteristics can be easily acquired with almost total transparency,

being less invasive than other methodologies. The latest advances in machine

learning have exposed the vulnerabilities of bot detectors [6,7], however at the

same time, these advances can be used to develop better bot detectors.

Our contributions with this work go a step forward in the bot detection

field incorporating behavioral modeling and improved learning methods based

on realistic synthetic samples (see Fig. 1):

Mouse Clicks

Mouse trajectory coordinates

Figure 2: An application example of our proposed mouse bot detection algorithm in com-

bination with a traditional image-based CAPTCHA. While the user completes the image

CAPTCHA task (cognitive challenge), our algorithm analyzes the mouse trajectories per-

formed during the task (neuromotor challenge).

•We propose two new methods for generating realistic mouse trajecto-

ries: i) a knowledge-based method based on heuristic functions, and ii)

a data-driven method based on Generative Adversarial Networks (GANs)

in which a Generator synthesizes human-like trajectories from a Gaussian

noise input. We demonstrate the usefulness of these synthetic trajectories

to train more accurate bot detectors. These Generators can be helpful in

many HCI research areas and applications.

•We propose BeCAPTCHA-Mouse, a new bot detector based on neuromo-

tor modeling of mouse trajectories and supervised classification trained

with human and synthetic data. As showed in Fig. 2, our proposed

mouse detection algorithm can be added in a transparent setup and en-

hance traditional CAPTCHAs based on cognitive challenges, for example

when you select the images in a visual CAPTCHA, or when you navigate

through a website.

•We present BeCAPTCHA-Mouse Benchmark6, the first public benchmark

for mouse-based bot detection including 10,000 human and synthetic tra-

jectories generated according to 10 different types of synthesized behaviors.

The inclusion of various types of synthetic samples (both for training and

testing BeCAPTCHA-Mouse) allows to train strong bot detectors. Also,

it allows comprehensive evaluations under various conditions including the

worst-case scenario in which bot attacks mimic human behavior using lat-

est machine learning advances. This benchmark can be helpful for other

HCI applications involving mouse dynamics beyond bot detection.

The rest of the paper is organized as follows. In Section 2we first discuss

the usage of mouse dynamics in the context of behavioral biometrics. Section

3describes the proposed methods for generating synthetic mouse trajectories.

Section 4presents our bot detector BeCAPTCHA-Mouse. Section 5describes

our experimental framework (BeCAPTCHA-Mouse Benchmark) and presents

the results obtained. Section 6compares our BeCAPTCHA-Mouse with re-

lated state-of-the-art CAPTCHA methods. Finally, Section 7summarizes the

conclusions and future works.

2. Mouse Dynamics in the Context of Behavioral Biometrics

Human-Machine interaction generates a heterogeneous flow of data from

multiple channels. This interaction generates patterns affected by: humans

(e.g. attitude, emotional state, neuromotor, and cognitive abilities), sensor

characteristics (e.g. ergonomics, precision), and task characteristics (e.g. easy of

use, design, usefulness). Modeling the user behavior using these heterogeneous

data streams is an ongoing challenge with applications in a variety of fields such

as security, e-health, gaming, or education [10,11,12]. Among this variety of

data sources, in the present paper we concentrate in behavioral biometric signals

[13].

6https://github.com/BiDAlab/BeCAPTCHA-Mouse

Uniq. Univ. Meas. Perf. Circ. Acce. Cog. Neu.

Keystroke ** ** *** *** ** ** ** ***

Stylometry * * * * * * *** *

Web-log ** * *** ** ** * * *** *

Mouse * ** *** *** * *** ** ***

Table 1: Biometric characteristics typically obtained in human-computer interaction. We

rate each factor with * (low), ** (medium), and *** (high). Uniq = Uniqueness, Univ =

Universality, Meas = Measurability, Perf = Performance, Circ = Circumvention, Acce =

Acceptability, Cog = Cognitive, Neu = Neuromotor.

The literature of behavioral biometrics in the context of Human-Computer

Interaction is large and includes several characteristics, e.g.: keystroking [14,15],

handwriting, touchscreen signals [16], stylometry [17,18], and mouse dynamics

[19,20]. Each characteristic has its pros and cons, therefore, a single biometric

characteristic is usually not suitable for all applications. The biometric research

community has identified several factors that determine the suitability of a bio-

metric characteristic to be used in a certain application [8].

Table 1rates these factors for biometrics characteristics typically obtained

from Human-Computer Interaction highlighting Mouse Dynamics, the focus in

the present paper. Note that we added two factors related to the nature of

the patterns obtained from these characteristics (Cognitive and Neuromotor

patterns) with respect to the characteristics defined by [8].

Now focusing in mouse dynamics for biometrics, in [19,20] researchers ex-

plored characteristics obtained from mouse tasks for user recognition. They

analyzed up to 68 global features (e.g. duration, curvature, mean velocity)

from mouse dynamics extracted during login sessions. Their results achieve up

to 95% authentication accuracy for passwords with 15 digits. Besides, mouse

dynamics can be combined with keystroke biometrics for continuous authenti-

cation schemes [21]. The fusion of both biometric modalities has been shown to

outperform significantly each individual modality achieving up to 98% authen-

tication accuracy [22,23]. In [24], the authors applied the Sigma-Lognormal

Model based on the Kinematic Theory [25] to compress mouse trajectories.

They suggested that mouse movements are the result of complex human motor

control behaviors that can be decomposed in a sum of primal movements. In

addition, in [26], the authors studied the relationship between eye gaze position

and mouse cursor position on a computer screen during web browsing and sug-

gested that there are regular patterns of eye/mouse movements associated to

the motor cortex system.

3. Mouse Trajectory Synthesis: Proposed Methods

In the present paper, a mouse movement is defined by the spatial trajectory

across time between two consecutive clicks, i.e., a sequence of points {x,y},

where x= [x1, . . . , xM], y= [y1, . . . , yM], and Mis the number of time samples.

We propose two methods for synthetically generating such mouse. A mouse

trajectory is defined by two main characteristics: the shape and the velocity

profile. In order to generate realistic synthetic samples, both characteristics

must be considered in the generation method.

3.1. Method 1: Knowledge-based Trajectories

We generate mouse tra jectories according to three different trajectory shapes

(linear, quadratic, and exponential) and three different velocity profiles (con-

stant, logarithmic, and Gaussian). We can synthesize many different mouse

trajectories that mimic human movements by varying the parameters of each

function. To generate a synthetic trajectory {ˆx,ˆy}with Mpoints, first we de-

fine the initial point [ˆx1, ˆy1] and ending point [ˆxM, ˆyM]. Second, we select one of

three velocity profiles: i) constant velocity, where the distance between adjacent

points is constant; ii) logarithmic velocity, where the distances are gradually in-

creasing (acceleration); and iii) Gaussian velocity, in which the distances first

increase and then decrease when they get close to the end of the trajectory (ac-

celeration and deceleration). Third, we generate a sequence ˆx between ˆx1and

ˆxMspaced according to the selected velocity profile. The ˆy sequence is then

Figure 3: Examples of mouse trajectories and their velocity profiles employed in this work: Ais

a real one extracted from a task of the database; Band Care synthetic trajectories generated

with the GAN network; D,Eand Fare generated with the knowledge-based approach. Note

that for each velocity profile (D= Gaussian, E= constant, F= logarithmic), we include the

three knowledge-based trajectories (linear, quadratic, and exponential).

generated according to the shape function. For example, for a shape defined by

the quadratic function ˆy=aˆx2+bˆx+c, we fit band cfor a fixed value of aby

using the initial and ending points. We repeat the process fixing either bor c.

The range of the parameters {a,b,c}explored is determined by analyzing real

mouse movements fitted to quadratic functions. Linear and exponential shapes

are generated similarly.

Fig. 3(trajectories D,E, and F) shows some examples of these mouse

trajectories synthesized. That figure also shows the 3 different velocity profiles

considered: the 3 trajectories in Ehave constant velocity, Fshows acceleration

(the distance between adjacent samples increases gradually), and Dhas initial

acceleration and final deceleration. We can generate infinite mouse trajectories

with this approach by varying the parameters of each function.

An important factor when synthetizing mouse trajectories is the number of

points (M) of the trajectory. This usually varies depending not only on the

Figure 4: The proposed architecture to train a GAN Generator of synthetic mouse trajectories.

The Generator learns the human features of the mouse trajectories and generate human-like

ones from Gaussian Noise.

length of the trajectory, but also on the direction, because different muscles

are involved when we perform mouse tra jectories in different directions. To

emulate this phenomenon, we calculate the mean and standard deviation of the

number of points for each of the 8 mouse trajectories from the human data used

in the experiments. Then, we synthetize trajectories with different number of

points following a Gaussian distribution with the calculated mean and standard

deviation.

3.2. Method 2: GAN-based Trajectories

For this approach we employ a GAN (Generative Adversarial Network) [27],

in which two neuronal networks, commonly named Generator and Discrimina-

tor, are trained one against the other (thus the adversarial). The architecture of

the GAN is depicted in Fig. 4. The aim of the Generator is to fool the Discrim-

inator by generating fake samples (mouse trajectories in this work) very similar

to the real ones while the Discriminator has to predict whether the sample comes

from the real set or is a fake created by the Generator. Once the Generator is

trained this way, then we can use it to synthesize mouse trajectories very similar

to the human ones.

The topology employed in both Generator and Discriminator consist of two

LSTM (Long Short-Term Memory) layers followed by a dense layer, very similar

to a recurrent auto-encoder. The dense layer of the Discriminator is used as a

classification layer to distinguish between fake and real mouse trajectories, while

the Generator employs the dense layer to build synthetic mouse trajectories.

Fig. 3shows two examples (tra jectories Band C) of synthetic mouse trajectories

generated with the GAN network and the comparison with a real one. We can

observe high similarity between the two synthetic examples and the real one.

Human mouse patterns such us the initial acceleration and the final trajectory

fine correction that we discussed before are automatically learned by the GAN

network and reproduced in the synthetic trajectories generated.

4. BeCAPTCHA-Mouse: Bot Detection based on Mouse Dynamics

The mouse is a very common device and its usage is ubiquitous in human-

computer interfaces. Bot detection based on mouse dynamics can be therefore

applied either in active or passive detectors.

In our BeCAPTCHA-Mouse bot detector we use mouse dynamics to extract

neuromotor features capable to distinguish human behavior from bots (see Fig.

1). Mouse dynamics are rich in patterns capable of describing neuromotor ca-

pacities of the users. Note that we do not claim to replace other approaches

(e.g. Google’s reCAPTCHA) by mouse-based bot detection, our purpose is to

enhance them by exploiting the ancillary information provided by mouse dy-

namics (see Fig. ??).

Our proposed method for bot detection consists in characterizing each mouse

trajectory with a fixed-size feature vector followed by a standard classifier. Each

trajectory characterized in this way can be classified individually using standard

classifiers into human or bot based on supervised training using a development

groundtruth dataset. When multiple trajectories are available, standard infor-

mation fusion techniques can be applied [16]. The more realistic the synthetic

data used as groundtruth for training the classifier the stronger the classifier.

In our experimental work we will use Support Vector Machine classifiers,

but any other standard classifier can be applied as well. The contribution and

Parameter Description

DiInput pulse: covered distance

t0iInitialization time: displacement in the time axis

µiLog-temporal delay

σiImpulse response time of the neuromotor system

θsi Starting angle of the stroke

θei Ending angle of the stroke

Table 2: Sigma-Lognormal features description.

success of our BeCAPTCHA-Mouse bot detector is not in the particular classifier

used, but in two other fronts (see Fig. 1): the high realism of the groundtruth

data used for training our classifiers (with the methods presented in Section 3),

and our proposed trajectory modeling using neuromotor features.

4.1. Neuromotor Analysis of Mouse Trajectories

By looking at typical mouse movements (see Fig. 5.a), we can observe some

aspects typically performed by humans during mouse trajectories execution: an

initial acceleration and final deceleration performed by the antagonist (activate

the movement) and agonist muscles (opposing joint torque) [28], and a fine-

correction in the direction at the end of the trajectory when the mouse cursor

gets close to the click button (characterized by a low velocity that serves to

improve the precision of the movement). These aspects motivated us to use

neuromotor analysis to find distinctive features in human mouse movements.

Neuromotor-fine skills, that are unique of human beings are difficult to emulate

for bots and could provide distinctive features in order to tell humans and bots

apart.

For this, we propose to model the trajectories according to the Sigma-

Lognormal model [29] from the kinematic theory of rapid human movements

[25]. The model states that the velocity profile of the human hand movements

(mouse movements in this work) can be decomposed into primitive strokes with

a Lognormal shape that describes well the nature of the hand movements ruled

by the motor cortex. The velocity profile of these strokes is modeled as:

|~vi(t)|=Di

√2πσi(t−t0i)exp (ln (t−t0i)−µi)2

−2σ2

i!(1)

where the parameters are described in Table 2. The velocity profile of the entire

hand movement is calculated as the sum of all these individual strokes:

~vr(t) =

i=1

~vi(t) (2)

where Nis the number of velocity strokes considered in the model. A complex

action like handwriting signature or mouse movements, is a summation of these

lognormals, each one characterized by the six parameters in Table 2. An ex-

ample of this is shown in Fig. 5.b, where the blue line is the velocity profile

|~v (t)|of the above human mouse task (Fig. 5.a), which is used as the input of

the Sigma-Lognormal model. The green dashed lines correspond to the individ-

ual lognormal signals |~vi(t)|generated as in [30], which describes a method to

automatically estimate both Nand the parameters in Table 2from an input

trajectory |~v (t)|. Finally, the red dotted line |~vr(t)|is the reconstruction of the

original velocity profile by summing all these generated individual lognormal

signals. We can observe that the reconstructed signal matches almost perfectly

with the original velocity profile of the human mouse movement, suggesting the

potential of the Sigma-Lognormal model to describe neuromotor mouse move-

ments. Lognormals with a high amplitude are typically observed during the

first part of the movement (agonist and antagonist activations), while smaller

lognormals occur during the fine correction. The differences in lognormal sizes

provide us information about the length of the trajectory (long trajectories have

usually larger velocities).

The neuromotor feature set proposed for bot detection is computed from the

six lognormal parameters described in Table 2. Each mouse trajectory generates

Nlognormal signals and each lognormal generates those 6 parameters from

Table 2. For each parameter, we calculate 6 features: maximum, minimum,

and mean for both halves of the trajectory. This is done because in natural

mouse movements the lognormal parameters are usually very different between

Figure 5: a) Example of the mouse task determined by 8 keypoints: the crosses represent

the keypoint where the user must click, red circles are the (x,y) coordinates obtained from

the mouse device, and the blue line is the mouse trajectory. b) and c) are examples of the

Lognormal decomposition of a human mouse movement and a synthetic linear tra jectory

respectively.

both halves of a given trajectory (e.g. Fig. 5.b). Additionally, we added the

number of lognormals Nthat each mouse trajectory generates as an additional

feature. This additional feature measures the complexity of the trajectory [31],

having many lognormals means that the mouse trajectory has many changes in

the velocity profile while few of them usually indicates more basic trajectories.

As a result, the neuromotor feature set has size 37.

5. Experiments

5.1. BeCAPTCHA-Mouse Benchmark: Database

The human mouse trajectories employed in this work were extracted from

Chao et al. [32] database, which is comprised of more than 200K mouse tra-

jectories acquired from 58 users who completed 300 repetitions of the task.

Acquisition of data fro meach subject took between 30 days and 90 days. In

each repetition, the task was to click 8 buttons that appeared in the screen

sequentially. This task was repeated twice in each session. Fig. 5.a shows an

example of the whole mouse movement task. Note that the buttons are placed

in a particular order to generate mouse trajectories with different directions

(rightwards, upwards, downwards, and oblique) and different lengths.

In the present work, we define a mouse trajectory as the mouse displacement

that occurs between two click buttons. Therefore, the mouse movement task of

Fig. 5.a is composed of 8 mouse trajectories. The raw data recorded during the

acquisition process was: the mouse position over the screen (x,yaxis position

in pixels), the event (movement or click), and timestamp of the event. The

experiments presented in this work are performed using a subset of the database

including 35 samples (randomly chosen) from each of the 58 users available

(more than 2K trajectories in total).

Fig. 5.c shows the decomposition of a synthetic knowledge-based trajectory

with linear shape. We can observe the huge differences between both lognormal

decompositions (the human trajectory and the synthetic one) by looking at the

shape of the lognormal signals. The synthetic trajectory has wider lognormals

and they are more symmetric than the human ones. Note that the Sigma-

Lognormal algorithm introduces a low-pass filter to the input signal, that is

the reason why the velocity profile of the synthetic trajectory (Fig. 5.c) is a

bit smoothed, but the difference between both synthetic and human velocity

profiles is still patent.

5.2. Experimental Protocol

We have extracted the proposed neuromotor features from human and syn-

thetic mouse trajectories (10K trajectories between both groups). We use an

SVM (Support Vector Machine) classifier with a RBF (Radial Basis Function)

kernel because of its good general performance in binary classification tasks.

The experiments are divided according to the 8 real mouse trajectories present

in the whole task. This means that we classify at trajectory level (i.e. the

mouse trajectory performed between two consecutive click buttons) instead of

classifying the whole task. This is because the task was designed to take into

account different directions and length trajectories, and therefore, different mus-

cles configurations are involved in each trajectory. In this way, we can analyze

which mouse trajectories are better to discriminate between humans and bots.

We train 10 different SVMs (one for each type of attack, see columns in Ta-

ble 3) using both human and synthetic trajectories. For each SVM, we train

the classifier by using 70% of both positive and negative samples and test with

the remaining 30% (randomly chosen), each experiment was repeated 5 times

and error rates were computed as the average of the 5 iterations.

The GAN network was trained using 60% of the human mouse trajectories in

the database. Training details: learning rate α= 2×10−4, Adam optimizer with

β1= 0.5, β2= 0.999, = 10−8, 50 epochs with a batch size of 128 samples for

both Generator and Discriminator. The loss function was binary crossentropy

for the Discriminator and mean square error for the Generator. The model was

trained and tested using Keras-Tensorflow.

Bot: Knowledge-Based

Linear Quadratic LogarithmicTrajectories

VP = 1 VP = 2 VP = 3 VP = 1 VP = 2 VP = 3 VP = 1 VP = 2 VP = 3

Bot:

GAN

8→1 6.7 8.3 3.3 18.7 10.0 6.0 17.3 12.4 8.04 1.8

1→2 1.1 5.6 2.5 6.1 5.6 8.3 8.3 3.4 10.0 3.7

2→3 1.1 0.8 3.9 7.2 1.7 15.7 9.4 3.9 11.1 1.3

3→4 1.7 2.2 6.7 5.0 2.2 13.9 5.0 2.3 12.8 0.3

4→5 2.2 3.9 2.5 7.8 2.2 12.8 7.2 3.4 13.3 2.5

5→6 1.7 4.4 6.1 3.9 1.1 15.0 3.9 5.7 11.1 1.5

6→7 5.0 4.4 3.3 12.2 8.9 8.9 15.0 10.3 10.6 1.5

Individual trajectories

7→8 4.9 7.2 7.2 10.6 11.1 9.1 13.3 16.1 17.7 0.8

Ours [Neuromotor] 2.3 2.9 4.1 6.1 6.5 7.7 6.4 7.6 7.9 3.9

Baseline [33]0.1 0.2 0.2 5.5 5.8 3.8 2.7 3.4 3.1 2.5

All

Ours [Neuromotor]+[33] 0.2 0.4 0.31.5 1.2 1.2 1.1 0.8 1.0 2.2

Table 3: Equal Error Rate (%) in the binary classification between each of the 8 human

trajectories and the synthetic ones. VP (Velocity Profile): VP = 1 constant velocity, VP = 2

initial acceleration, VP = 3 initial acceleration and final deceleration.

5.3. BeCAPTCHA-Mouse Benchmark: Results

Table 3shows the final results for all classification schemes. The first 8 rows

present the 8 trajectories derived from the movements between the 8 keypoints

(plotted in Fig. 5.a). The table shows the classification errors in % (human vs

bot) for the different synthetic trajectories (in columns) generated in this work.

The results are presented in terms of EER (Equal Error Rate) defined as the

point where the False Positive Rate and the False Negative Rate are equal.

First, comparing among the different trajectories, we can observe that the

shorter ones (8 →1, 6 →7, and 7 →8) show higher classification errors

compared to the larger ones. Short trajectories generate less neuromotor in-

formation: initial acceleration, final deceleration, and trajectory corrections are

less pronounced in short trajectories. Second, logarithmic trajectory shapes

achieve the worst classification performance, as we expected, because the shape

of logarithmic functions fit better the human trajectories shapes. Third, the

most significant parameter when synthetizing trajectories is the velocity profile.

When VP = 3 (i.e., initial acceleration and final deceleration), the synthetic

trajectories are able to fool the classifier up to 17% of the times. This confirms

that the velocity profile of human mouse trajectories plays and important role

when describing human features in mouse dynamics. Four, the GAN Generator

(last column in Table 3) results in lower classification errors compared with the

knowledge-based method. This is surprising after visualizing the high similarity

between human and GAN-generated trajectories (see Fig. 3Avs Band Avs

C). We interpret this result with care: on the one hand it demonstrates that

out bot detection approach is powerful against realistic and sophisticate fakes,

but on the other hand the GAN Generator can be improved to better fool our

detector.

The last three rows in Table 3present the results when features from all 8

trajectories are combined (each SVM is trained using features from all 8 tra-

jectories). Additionally, we compare the performance achieved with existing

approaches [33]. The feature set proposed in [33] consists of 6 global features:

duration, distance, displacement, average angle, average velocity, and move ef-

ficiency (distance over displacement). The results suggest that the feature set

proposed in [33] outperforms the neuromotor features proposed here only for

Linear synthetic trajectories. The best performance is obtained overall with an

extended set composed by both sets of features. The extended set has the best

results with an average around 1% EER independently of the type of synthetic

trajectory.

Finally, Table 4shows the EER when all types of attacks are used to train

and test the system. In this case, one SVM is trained using trajectories from all

8 directions and synthetic samples from all 10 types of attacks. The results show

that the neuromotor feature set allows to reduce the error by 36% in comparison

with the previous existing method [33]. These results demonstrate the potential

of mouse dynamics features to distinguish between human and synthetic mouse

movements. Additionally, we show the performance of a one class SVM classifier

trained using only real samples. As can be seen, the classifier trained only with

real samples was not capable to detect most of the attacks with error rates

Features Training

Only Real Real+Fake

Baseline [33] 34.7% 4.4%

Ours [Only Neuromotor] 35.6% 10.2%

Ours [Neuromotor + Baseline] 40.1% 2.8%

Error Reduction ↑15% ↓36%

Table 4: Equal Error Rate (%) in bot detection of the different feature sets for models trained

with and without synthetic samples (fakes) and evaluated using human samples and fake

samples. The last row shows the error reduction compared to the set proposed in [33].

over 34% either for baseline set and neuromotor features. The importance of

synthetic samples is twofold: i) evaluation of bot detection algorithms under

challenging attacks generated according to different methods; and ii) training

better detectors to model both human and synthetic behaviors. The results in

Table 4show the potential of the synthetic samples and its usefulness to train

better models capable to deal with all types of attacks.

6. BeCAPTCHA and Complementarity with the State of the Art

BeCAPTCHA-Mouse is a bot detector based on the behavior modeling of

human-machine interaction. The exploitation of behavioral biometrics for bot

detection is an open research line with large opportunities and challenges. These

challenges include the study of new ways of interactions such as keystroke or

touch [34], the applications to mobile scenarios, or the circumvention to attacks.

We want to highlight that behavioral CAPTCHAs are compatible with previous

CAPTCHA technologies and it could be added as a new cue to improve existing

bot detection schemes in a multiple classifier combination [16] (see Fig. 6).

Table 5shows some of the main features of different existing CAPTCHA

methods. As we commented in the introduction section, most of them have

been defeated by machine learning algorithms. In fact, the last version of the

Google CAPTCHA, named reCAPTCHAv3, that measures mouse dynamics

ReCAPTCHA

Context

BeCAPTCHA

Human-Computer

Interaction

Cognitive and Behavior

Analysis

Combination of

Experts

Fusion

Decision

Human

Bot

Data

IP, MAC, …

Mouse

Figure 6: Block diagram of multimodal bot detection. The response of the bot detector is a

combination of responses from different experts. The bot detector proposed in this work can

be used independently or in combination with existing bot detectors.

and web browsing interactions between the user and the web site to decide

whether the user is a bot or not, was recently hacked in [35] by synthetizing

mouse trajectories using reinforcement learning techniques. The main problem

of these CAPTCHA methods is that they only measure cognitive human skills

(e.g. character recognition from distorted images, class-objects identification in

a set of images, or speech translation from distorted audios). Trying to ensure a

very accurate bot detection makes these CAPTCHAs difficult to perform even

for humans. The main goal of our proposed method is to focus more on human

behavioral skills rather than on cognitive ones. Neuromotor skills reveal human

features useful for bot detection just with simple mouse trajectories. To the

best of our knowledge, there are only a very limited number of works using

mouse biometrics for bot detection. The most related to our research are [33]

and [35]. In [35] they synthetize mouse trajectories over a grid to hack the

Google reCAPTCHA v3 algorithm, and in [33] they extract global features (e.g.

duration, average speed, displacement) from mouse and keystroke patterns to

conduct a case study in the detection of blog bots for online blogging systems.

While previous work in mouse dynamics ([19,20,33]) focused on basic cues like

duration or average speed, in this work we go a step forward by focusing on

the analysis of entire mouse trajectories, using the Sigma-Lognormal model to

extract human features that characterizes better human behaviors.

Method Cog. Beh. Usability Security

Audio CAPTCHA *** * * *

Image CAPTCHA *** * * *

Text CAPTCHA *** * * *

reCAPTCHA v3 * ** *** **

Our method ** ** *** ***

Table 5: Characteristics of several CAPTCHA methods. We rate each factor as low (*),

medium (**) and high (***). Cog = Cognitive, Beh = Behavioral.

7. Conclusions and Future work

We have explored behavioral biometrics for bot detection during human-

computer interaction. In particular, we have analyzed the capacity of mouse

dynamics to describe human neuromotor features. Our conclusions in compari-

son to state-of-the-art works suggest that there is unexploited potential of mouse

dynamics as a behavioral biometric for tasks such as bot detection.

In concrete, we have proposed BeCAPTCHA-Mouse, a bot detection algo-

rithm based on mouse dynamics, and a related benchmark7, the first one public

for research in bot detection and other mouse-based research areas including

HCI, security, and human behavior.

Additionally, we have proposed and studied two new methods for generat-

ing synthetic mouse trajectories of varying level of realism. These generators

are very useful both training stronger bot detectors, and evaluating them in

comprehensive and worst case scenarios. These generators are also valuable for

related research problems beyond bot detection involving mouse dynamics.

In our experiments we have observed the main features of human mouse

trajectories (e.g. initial acceleration, final deceleration, and fine trajectory cor-

rection). Based on that we have developed a neuromotor feature representation

using the Sigma-Lognormal model [25,29]. Using the proposed neuromotor fea-

7https://github.com/BiDAlab/BeCAPTCHA-Mouse

ture representation and training standard classifiers making use of the proposed

synthetic mouse trajectories, we have been able to discriminate between humans

and bots with up to 93% of accuracy, even with bots of high realism, and only

one mouse trajectory as input (between two consecutive clicks). This proves the

potential of mouse dynamics for Turing tests.

As future work, we aim at improving the neuromotor feature set by calculat-

ing secondary features inferred from the main ones. Also, we propose to combine

both synthesis methods by using the knowledge-based trajectories as the input

of the GAN model instead of Gaussian noise. This technique could generate

more sophisticate and human-like trajectories. Finally, in this paper we only

considered mouse trajectories acquired from mouse devices. We also propose

to analyze mouse-pad trajectories normally performed when using laptops as

another line of research.

Acknowledgements

This work has been supported by projects: IDEA-FAST (H2020-IMI2-2018-

853981), PRIMA (H2020-ITN-2019-860315), TRESPASS-ETN (H2020-ITN-2019-

860813), BIBECA (RTI2018-101248-B-I00 MINECO/FEDER), and BioGuard

(Ayudas Fundacin BBVA a Equipos de Investigacin Cientfica 2017). Spanish

Patent Application P202030066.

References

[1] A. M. Turing, Computing Machinery and Intelligence, Springer Nether-

lands, Dordrecht, 2009, pp. 23–65.

[2] A. P. Saygin, I. Cicekli, V. Akman, Turing test: 50 years later, Minds and

Machines 10 (2000) 463–518.

[3] J. Svilpis, The science-fiction prehistory of the turing test, Science Fiction

Studies 35 (3) (2008) 430–449.

[4] Y. Gorodnichenko, T. Pham, O. Talavera, Social media, sentiment and

public opinions: Evidence from #Brexit and #USElection, Working Paper

24631, National Bureau of Economic Research (2018).

[5] Y. Xie, S.-Z. Yu, A large-scale hidden semi-markov model for anomaly

detection on user browsing behaviors, IEEE/ACM Trans. Netw. 17 (2009)

54–65.

[6] E. Bursztein, M. Martin, J. Mitchell, Text-based CAPTCHA strengths and

weaknesses, in: Proceedings of the 18th ACM Conference on Computer and

Communications Security, CCS 11, Association for Computing Machinery,

New York, NY, USA, 2011, pp. 125–138.

[7] K. Bock, D. Patel, G. Hughey, D. Levin, unCaptcha: A low-resource defeat

of reCaptcha’s audio challenge, in: 11th USENIX Workshop on Offensive

Technologies (WOOT 17), USENIX Association, Vancouver, BC, 2017.

[8] A. K. Jain, K. Nandakumar, A. Ross, 50 years of biometric research: Ac-

complishments, challenges, and opportunities, Pattern Recognition Letters

79 (2016) 80–105.

[9] J. Fierrez, A. Pozo, M. Martinez-Diaz, J. Galbally, A. Morales, Bench-

marking touchscreen biometrics for mobile authentication, IEEE Trans. on

Information Forensics and Security 13 (11) (2018) 2720–2733.

[10] H. J. Escalante, H. Kaya, A. A. Salah, S. Escalera, Y. G;ltrk, U. Gl, X. Bar,

I. Guyon, J. C. S. Jacques, M. Madadi, S. Ayache, E. Viegas, F. Gurpinar,

A. S. Wicaksana, C. Liem, M. A. J. Van Gerven, R. Van Lier, Model-

ing, recognizing, and explaining apparent personality from videos, IEEE

Transactions on Affective Computing.

[11] E. Nosakhare, R. Picard, Toward assessing and recommending combina-

tions of behaviors for improving health and well-being, ACM Trans. Com-

put. Healthcare 1 (1).

[12] H. Shrobe, D. L. Shrier, A. Pentland, Behavioral biometrics, in: New So-

lutions for Cybersecurity, 2018, pp. 365–377.

[13] A. Alzubaidi, J. Kalita, Authentication of smartphone users using behav-

ioral biometrics, IEEE Communications Surveys Tutorials 18 (3) (2016)

1998–2026.

[14] C. C. Tappert, S.-H. Cha, M. Villani, R. S. Zack, A keystroke biometric

system for long-text input, International Journal of Information Security

and Privacy 4 (1) (2010) 32–60.

[15] A. Acien, J. V. Monaco, A. Morales, R. Vera-Rodriguez, J. Fierrez, Type-

net: Scaling up keystroke biometrics, arXiv:2004.03627.

[16] J. Fierrez, A. Morales, R. Vera-Rodriguez, D. Camacho, Multiple classifiers

in biometrics. part 2: Trends and challenges, Information Fusion 44 (2018)

103–112.

[17] J. C. Stewart, J. V. Monaco, S. Cha, C. C. Tappert, An investigation

of keystroke and stylometry traits for authenticating online test takers,

in: Proceedings of the 2011 International Joint Conference on Biometrics

(IJCB), 2011, pp. 1–7.

[18] H. Locklear, S. Govindarajan, Z. Sitov, A. Goodkind, D. G. Brizan,

A. Rosenberg, V. V. Phoha, P. Gasti, K. S. Balagani, Continuous authen-

tication with cognition-centric text production and revision features, in:

Proceedings of the IEEE International Joint Conference on Biometrics,

2014, pp. 1–8.

[19] A. A. E. Ahmed, I. Traore, A new biometric technology based on mouse

dynamics, IEEE Transactions on Dependable and Secure Computing 4 (3)

(2007) 165–179.

[20] H. Gamboa, A. L. N. Fred, A. K. Jain, Webbiometrics: User verification

via web interaction, in: Proceedings of the 2007 Biometrics Symposium,

2007, pp. 1–6.

[21] T. Sim, S. Zhang, R. Janakiraman, S. Kumar, Continuous verification us-

ing multimodal biometrics, IEEE Transactions on Pattern Analysis and

Machine Intelligence 29 (4) (2007) 687–700.

[22] S. Mondal, P. Bours, A study on continuous authentication using a com-

bination of keystroke and mouse biometrics, Neurocomputing 230 (2017)

1–22.

[23] K. O. Bailey, J. S. Okolica, G. L. Peterson, User identification and authen-

tication using multi-modal behavioral biometrics, Computers & Security

43 (2014) 77–89.

[24] D. Mart´ın-Albo, L. A. Leiva, J. Huang, R. Plamondon, Strokes of insight,

Inf. Process. Manage. 52 (6) (2016) 989–1003.

[25] R. Plamondon, A kinematic theory of rapid human movements, Biological

Cybernetics 72 (4) (1995) 295–30.

[26] M. C. Chen, J. R. Anderson, M. H. Sohn, What can a mouse cursor tell us

more? correlation of eye/mouse movements on web browsing, in: Proceed-

ings of the CHI 01 Extended Abstracts on Human Factors in Computing

Systems, CHI EA 01, New York, NY, USA, 2001, pp. 281–282.

[27] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,

S. Ozair, A. Courville, Y. Bengio, Generative adversarial nets, in: Proceed-

ings of the 27th International Conference on Neural Information Processing

Systems - Volume 2, Cambridge, MA, USA, 2014, pp. 2672–2680.

[28] M. Djioua, R. Plamondon, A new algorithm and system for the charac-

terization of handwriting strokes with delta-lognormal parameters, IEEE

Transactions on Pattern Analysis and Machine Intelligence 31 (11) (2009)

2060–2072.

[29] A. Fischer, R. Plamondon, Signature verification based on the kinematic

theory of rapid human movements, IEEE Transactions on Human-Machine

Systems 47 (2) (2017) 169–180.

[30] O. Loyola-Gonz´alez, R. Monroy, M. A. Medina-P´erez, B. Cervantes, J. E.

Grimaldo-Tijerina, An approach based on contrast patterns for bot de-

tection on web log files, in: I. Batyrshin, M. d. L. Mart´ınez-Villase˜nor,

H. E. Ponce Espinosa (Eds.), Advances in Soft Computing, Springer Inter-

national Publishing, Cham, 2018, pp. 276–285.

[31] R. Vera-Rodriguez, R. Tolosana, J. Hernandez-Ortega, A. Acien,

A. Morales, J. Fierrez, J. Ortega-Garcia, Modeling the complexity of sig-

nature and touch-screen biometrics using the lognormality principle, in:

R. Plamondon, A. Marcelli, M. A. Ferrer (Eds.), The Lognormality Prin-

ciple and its Applications, World Scientific, 2019.

[32] C. Shen, Z. Cai, X. Guan, R. Maxion, Performance evaluation of anomaly-

detection algorithms for mouse dynamics, Computers & Security 45 (2014)

156–171.

[33] Z. Chu, S. Gianvecchio, H. Wang, Bot or Human? A Behavior-Based Online

Bot Detection System, Springer International Publishing, Cham, 2018, pp.

432–449.

[34] A. Acien, A. Morales, J. Fierrez, R. Vera-Rodriguez, I. Bartolome, Be-

CAPTCHA: detecting human behavior in smartphone interaction using

multiple inbuilt sensors, in: Proceedings of the AAAI Workshop on Artifi-

cial for Cyber Security (AICS), 2020.

[35] I. Akrout, A. Feriani, Akrout, Hacking google reCAPTCHA v3 using rein-

forcement learning, in: Conference on Reinforcement Learning and Decision

Making, 2019.